Immersive 3-D Video Conferencing: Challenges, Concepts, and Implementations

Transcription

1 Proc. SPIE Visual Communications and Image Processing (VCIP), Lugano, Switzerland, July Immersive 3-D Video Conferencing: Challenges, Concepts, and Implementations Peter Eisert Image Processing Department Fraunhofer Institute for Telecommunications, Heinrich-Hertz-Institute Einsteinufer 37, D Berlin, Germany URL: eisert ABSTRACT In this paper, a next generation 3-D video conferencing system is presented that provides immersive telepresence and natural representation of all participants in a shared virtual meeting space. The system is based on the principle of a shared virtual table environment which guarantees correct eye contact and gesture reproduction and enhances the quality of human-centered communication. The virtual environment is modeled in MPEG-4 which also allows the seamless integration of explicit 3-D head models for a low-bandwidth connection to mobile users. In this case, facial expression and motion information is transmitted instead of video streams resulting in bit-rates of a few kbit/s per participant. Beside low bit-rates, the model-based approach enables new possibilities for image enhancements like digital make-up, digital dressing, or modification of scene lighting. Keywords: Immersive Video Conferencing, Tele-Collaboration, Shared Virtual Table Environment, Model- Based Coding, View Synthesis 1. INTRODUCTION The idea of video conferencing and video telephony has fascinated researchers for a long time and new devices have been developed as soon as new technologies became available. Already in 1927, a video phone conversation with life video transmission over telephone lines from Washington D.C. to an auditorium in Manhattan was demonstrated. The Bell picturephone shown in Fig. 1 used a mechanical Nipkow Disk for sampling the images and had a display of size 2 by 2.5 inches. Figure 1. Bell s picturephone system of Over time, devices got smaller while display sizes increased. In the sixties, a couple of different video phones were developed which enabled video transmission over telephone lines. The left hand side of Fig. 2 shows an example of such a picture phone. Unfortunately, these technologies never became well accepted due to costs, limited availability, and restricted gain in communication. First with the introduction of digital computers, cheap and powerful hardware and software devices have been developed and, today, video conferencing is rapidly advancing. Especially for decision-making in a wide range of applications, from world-wide operating companies, international research projects to global investment banking, the interest in tele-conferencing has increased in recent years.

2 Figure 2. Left: Picturephone from Right: Hydra system. Nevertheless, video conferencing today is limited in its support of natural human-centered communication. Body postures and subtle movements, gaze direction, room acoustics, and joint interactions are often misrepresented, lacking, or wholly absent. Eye contact is still a problem and also verbal and non-verbal communication between other participants can often not be followed. To address these limitations, several experimental systems have been proposed in the last few years. The Hydra system depicted on the right hand side of Fig. 2, e.g., uses three independent communication devices with three displays distributed on a desk to ensure a spatial and acoustical separation of the participants. Figure 3. Concept of an immersive shared virtual table video conferencing system. Beside the reproduction of spatial relations, large displays 1 and other benefits of tele-immersion are exploited in order to give the participants the impression of being present in a shared virtual table environment (SVTE) suggesting spatial proximity and enabling a higher degree of natural interaction and effective collaboration. In this sense, the participants should be able to make use of rich communication modalities as similar as possible to those used in a face-to-face meeting. The exact reproduction of gestures, eye contact, sound direction, and multiple simultaneous inter-participant conversations are examples of these requirements. The main objective is to enable conferees located in different geographical places to meet around a virtual table, appearing at each

3 station in such a way to create a convincing impression of presence as illustrated in Fig. 3. First approaches in this area were limited to strictly graphical environments using simple avatars for representing the remote conferees. A lot of systems have been proposed in this context during the last decade. One example for these research activities was the former Europe ACTS project COVEN that demonstrated the benefits of the SVTE concept by a networked business game VR application. 2 Later approaches were more and more focused on a seamless integration of 2-D video images or even 3-D video avatars into Collaborative Virtual Environments (CVE) to increase realism. 3, 4 Some of these approaches were driven by the MPEG-4 multimedia standard that offers new powerful coding and composition tools for such purposes. 5 Figure 4. Left: Setup of NTII at UNC. 6 Right: TELEPORT System. 7 An attractive SVTE approach is the one of tele-cubicles. 6 8 A common feature of these proposals is a special system set-up, where the participants are situated symmetrically around the shared table, with each conferee appearing on an own screen as illustrated in Fig. 4. Note that the symmetric geometry of this setup guarantees eye contact, gaze awareness and gesture reproduction. Thus, everybody in the session can observe who is talking to whom or who is pointing at what. In the next section, the Virtual Team User Environments (VIRTUE) system 9 is proposed that offers all benefits of the tele-cubicle approach and extends it towards an integration into shared virtual working spaces. A seamless transition between the real working desk in front of the display and the virtual conference table at the screen gives the user the impression of being part of an extended perception space. Moreover, the remote participants are rendered seamlessly und under correct perspective view into the virtual conference scene. Scene depth information obtained from stereo cameras and encoded and streamed using MPEG-4 10 enables view synthesis in a virtual 3-D scene. Replacing the depth maps by explicit 3-D computer graphics head-and-shoulder models leads to an enormous reduction in bit-rate for the transmission of scene data, since shape and texture data need not be updated for each frame. Such a model-based coding system, 11 which can be used as a fall-back solution for mobile devices connected to the Internet over low-bandwidth channels, is presented in Section VIRTUAL TEAM USER ENVIRONMENT In this section, we present an extension to the Virtual Team User Environments (VIRTUE) system 9, 12 that uses the state-of-the-art multimedia standard MPEG-4 for transmission and representation of scene and texture data. Due to usage of the MPEG-4 signaling protocol DMIF, a sophisticated 3-D video processing, and the MPEG-4 scene description language BIFS, the system architecture supports an arbitrary terminal setup, interoperability with other terminal configurations, and scalability in terms of the number of participants and it makes it open for future extensions and developments towards immersive tele-collaboration, multi-user terminals, and joint applications.

4 2.1. Shared Virtual Table Concept The basic idea of the shared virtual environment concept is to place 3-D video reproductions of a given number of participants at predefined positions in a shared virtual environment. For this purpose, the conferees are captured at each terminal by a multiple camera setup as shown in Fig. 5 and the desired 3-D video representation of the local person is extracted from the images. Figure 5. Left: Setup for a 3-party conference. Right: VIRTUE setup. Then, the 3-D video objects of all conferees are grouped virtually around the shared table. Ideally, this is done in an isotropic manner in order to obtain social and geometric symmetry. Hence, in the case of a 3-party conference, the participants form an equilateral triangle. In the case of four parties, it would be a square, and so on. Following such generic composition rules and knowing the number of participants, the same SVTE can be built at each terminal from previously loaded scene descriptions and the 3-D video streams. Figure 6. Rendering of the virtual 3-D conference scene. Based on this generic scene composition, individual views of the virtual conference environment can be rendered by using a virtual camera as illustrated in Fig. 6. Locally, the position of the virtual camera has to move coincidently with the current position of the conferee s head, which is permanently registered by a head tracker. Thus, supposing that the geometrical parameters of the multi-view capture device, the virtual scene, and the virtual camera are well fitted to each other, it is ensured that all conferees see the scene under the correct perspective view, even while changing their own viewing position. This geometrical coincidence provides all desired attributes mentioned in the introduction eye contact, gaze awareness, gesture reproduction, natural conference situation, and high amount of realism. In addition, the support of head motion parallax allows the conferees to change their viewing position in order to watch the scene from another perspective, to look behind objects, or to look at a previously occluded object.

5 2.2. Architecture of the System Fig. 7 outlines the system architecture of the 3-D video conferencing system. After multi-view capturing, the video frames are segmented to separate the person s silhouette from the background. As a result, the conferees are represented by arbitrarily shaped video objects, which can be integrated seamlessly into virtual environments. To extract depth information, disparity estimation is performed on the rectified video objects resulting in dense disparity maps. Both video objects and disparity maps are efficiently encoded using MPEG-4. Figure 7. Architecture of the3-d video conference system. The system concept takes advantage of several particular features of the MPEG-4 multimedia standard. MPEG-4 allows the encoding of arbitrarily shaped video objects and provides auxiliary alpha planes to transmit additional pixel information associated to the color data. These additional planes can be used for a joint transmission of disparity maps and video objects. After encoding, the packets are streamed to other participating terminals via RTP. Simultaneously, the terminal receives video streams from the other conference partners and decodes them with multiple MPEG-4 video decoders. The shaped video objects and audio data are synchronized to each other and integrated into the SVTE scene represented by the MPEG-4 scene description language BIFS. Finally, an MPEG-4 compositor is used for rendering purposes. This compositor is able to handle user events (head tracker or interaction with scene content) as well as scene updates sent by a server. The MPEG-4 video objects are three-dimensionally warped using image-based rendering techniques before they are integrated into the scene. Depending on the current input from the head tracker, the correct perspective view is calculated and the adapted view is inserted into the BIFS scene as a 2-D video Foreground Background Segmentation In the finally rendered scene, the participants appear as arbitrarily shaped video objects, seamlessly integrated into the virtual environment. This requires a segmentation of the moving person from the background which is assumed to remain rather static. Initially, the background is captured and a change detection scheme compares this reference image with the current video data and provides a segmentation mask. 5 The reference image is permanently updated to cope with slight changes of illumination or scene content. This baseline algorithm has been improved in speed and quality in order to meet the real-time constraints for full CCIR601 resolution and video frame rate. Its performance has further been improved by adding a fast and efficient shadow detection tool. 13 For a robust segmentation, this is particularly important, because shadows at the table can usually not be avoided, even under optimal illumination conditions. The effect of shadow detection on the segmentation result is shown in Fig. 8.

6 Figure 8. Left: Original frame. Middle: Segmented foreground without shadow detection. Right: with shadow detection Disparity Estimation and Depth Analysis To extract the required 3-D representation, the depth of the captured video object is analyzed by disparity matching on the basis of the rectified images. We have developed a new hybrid block- and pixel-recursive approach, which is able to compute the disparity fields in real-time on a state-of-the-art PC. 14 Apart from a considerable reduction of computational load, the algorithm leads to spatio-temporally consistent depth maps, particularly important for view synthesis, since temporal inconsistencies may cause annoying artifacts in the synthesized views. In order to deal with occlusions caused by hands, arms, or the head, several post-processing techniques are applied. In a first step, critical regions are detected by a consistency check between disparity fields for a left to right and a right to left match. Unreliable disparity values in occluded areas are filled efficiently by suitable extrapolation techniques. The segmentation of the hands and arms which cause severe depth discontinuities is further refined by tracking and segmenting the hands using motion and skin color information. Fig. 9 shows two rectified images, the disparity fields with occluded areas, and the final disparity map after post-processing. Figure and 2. Image: Rectified frames from the left and right camera. occluded areas. 5. Image: Disparity map after post-processing. 3. and 4. Image: Disparity fields with 2.5. Head Tracking The perspective view of the scene presented on the display depends on the viewer s position. This requires an accurate estimation of the viewer s position in the 3-D space, which is accomplished by the head-tracking module. The chosen approach is based on a skin color segmentation technique jointly with a facial feature tracker searching eye positions. Due to the restriction to video conference scenarios, several assumptions can be made to facilitate the tracking of the viewer s head. The obtained 2-D positions of the eyes in two images can then be used for an accurate calculation of the 3-D head position View Synthesis and Virtual Scene Composition The calculated head position and the analyzed depth of the transmitted video objects provide sufficient information to synthesize novel virtual views of the remote conferees. In order to meet real-time constraints 15, 16 and

7 efficient occlusion handling, a new view synthesis algorithm has been developed in the VIRTUE project which is able to take these issues into account. 17 Figure 10. Composed scene containing computer graphics and video objects. At the end of the processing chain, the virtual conference scene has to be composed and displayed onto the screen as depicted in Fig. 10. The virtual scene is represented by a number of polygons in 3-D space and encoded by the BIFS scene description language of MPEG-4. In this polygon-based scene representation the participants are substituted by 2-D rectangles positioned around the virtual table. In this sense, the synthesized novel views of the remote participants are treated as textures and are transferred directly to the graphics card D MODEL-BASED VIDEO CONFERENCING In the previous section, the participants are described by planar computer graphics objects. Onto these planes, the recorded video is projected after warping according to the estimated disparity maps. This has the advantage of very natural looking results, since the original video is exploited in the image-based view synthesis. On the other hand, video streams from each participant together with their corresponding disparity maps have to be transmitted to the client. Dependant on the number of participants, this results in rather high bit-rates even if the streams are encoded with MPEG-4. Current wireless channels are usually not able to handle such bit-rates which prevents the participation of users from mobile devices like PDAs or smart phones. Figure 11. Virtual Video Conferencing on a Mobile Device.

8 In this section, we propose an alternative model-based approach 11, 18, 19 that can be used as a fall-back solution for virtual conferencing systems which are connected to the Internet via low-bandwidth channels. Instead of transmitting texture and depth information of all conferees for each frame, 3-D models of the participants are created representing their appearance. The models are generated only once in an initial phase, encoded, and transmitted to all devices. During the video conference, motion and facial expressions of the conferees are tracked over time and only a small set of facial animation parameters has to be streamed over the network. Experiments have shown, that bit-rates of a few kbit/s per participant can be achieved which enables virtual video conferencing also on mobile devices as illustrated in Fig. 11. The decoder only has to move and deform the 3-D head-andshoulder models in the virtual conferencing scene and render it using computer graphics techniques. Rendering and the encoding and streaming of 3-D head models and facial animation parameters (FAPs) can be realized with MPEG-4 in a standard compliant way. 10 Thus, the compositor of the aforementioned immersive video conferencing system can incorporate these explicit models with only minor changes. The architecture of such a model-based coding component with image analysis and view synthesis is depicted in Fig. 12. Video Analysis Estimation of FAPs Coder Parameter entropy coding Channel about 1 kbit/s Parameter decoding Decoder Synthesis Animation and rendering of the head model Video Head Model Shape Texture Dynamics Figure 12. Model-based Video Codec. Besides the low bit-rate, the usage of explicit 3-D models for the people in the scene also has some other implications on the virtual conferencing system. A sophisticated real-time segmentation of the conferees can be avoided, since the computer models already provide an inherent segmentation. Moreover, view point transformations can easily be applied by simply moving the virtual camera, since all objects in the scene are represented by 3-D models and no image-based warping is required. Another problem in mobile virtual conferencing is an eventually unfavorable appearance or dressing of the conferee depended on his or her current location or situation. By using pre-stored 3-D models for transmission from the mobile device to the other participants, the desired look can be ensured even under awkward situations. This way, digital dressing and digital make-up can be easily applied Facial Expression Analysis The most challenging part of facial analysis and synthesis is the estimation of 3-D facial motion and deformation from two-dimensional images. Due to the loss of one dimension caused by the projection of the real world onto the image plane, this task can only be solved by exploiting additional knowledge of the objects in the scene. In particular, the way the objects move can often be restricted to a low number of degrees of freedom that can be described by a limited set of parameters. We have developed a new 3-D model-based method for the estimation of facial expressions that makes use of an explicit parameterized 3-D human head model describing shape, color, and motion constraints of an individual person. 20 This model information is jointly exploited with spatial and temporal intensity gradients of the images. Thus, the entire area of the image showing the object of interest is used instead of dealing with discrete feature points, resulting in a robust and highly accurate system for modelbased estimation of MPEG-4 facial animation parameters. A linear and computationally efficient algorithm can

9 be derived for different scenarios. The scheme is embedded in a hierarchical analysis-synthesis framework to avoid error accumulation in the long-term estimation Illumination Analysis For natural and uncontrolled video capture conditions, often occurring in mobile environments, scene lighting often varies over time. This illumination variability not only has a considerable influence on the visual appearance of the objects in the scene, but also on the performance of computer vision algorithms or video coding methods. The efficiency and robustness of these algorithms can be significantly improved by removing the undesired effects of changing illumination. In our facial analysis and synthesis system, we model also lighting properties in the scene and estimate them. 21 The explicit knowledge about light sources and reflection properties not only increases robustness of the facial expression analysis but also enables new possibilities for improvement or manipulation of lighting conditions. For example, variations in illumination can be removed in a post-processing step and the lighting conditions can be adjusted to the one in the virtual conference room. This is illustrated in Fig. 13, where the upper row refers to the original image sequence while the lower row depicts the corresponding frames after lighting change estimation and removal. Figure 13. Upper row: original sequence, lower row: corresponding illumination-compensated frames with constant lighting Model-based View Synthesis Synthesizing new views is accomplished by simply rendering the 3-D models in the virtual scene. MPEG-4 offers possibilities to describe head models as well as streaming of facial animation parameters. The color of the person is defined by a texture map which is created from the video frames. For the definition of the individual shape, multiple options are available in MPEG-4. We use our own generic triangle mesh which is adapted to the individual and defined using the scene description BIFS. A facial animation table (FAT) 10 specifies the mesh deformations and transforms which occur when changing a specific facial animation parameter. After the virtual scene, defined by the BIFS tree, is transmitted to all participants of the video conference, only facial animation parameters are sent for each frame after having been estimated from the camera images. Bit-rate measurements have shown, 20 that about 1 kbit/s is sufficient to stream these parameters. Fig. 14 shows an example for the proposed model-based video streaming system. The upper row depicts three frames from a head-and-shoulder video sequence of one participant in a conference. From this monocular image sequence, the 3-D head model shown in the middle row of Fig. 14 is created and facial expression analysis is performed. The estimated parameters are encoded and sent to the decoder which renders the models after

10 Figure 14. Upper row: sequence Peter, middle row: animated wireframe model, lower row: synthesized model frames generated by rendering the 3-D head model. applying local deformation and motion. The middle and lower row of Fig. 14 refer to the output of the decoder after rendering as a wire-frame or with a textured model. 4. CONCLUSIONS We have presented a new concept for immersive video conferencing using a shared virtual table environment (SVTE) which allows a very intensive communication and effective collaboration. It combines the benefits of former VR-based collaborative virtual environment approaches and video-based tele-cubicle systems with those of mixed reality applications. A seamless transition between the real and virtual conference table gives the user the impression of an extended perception space. The usage of 3-D video objects for the representation of remote conferees enables natural representation of gestures, eye contact and gaze. Due to the MPEG-4 multimedia standard, the system provides open and flexible system architectures. Beside arbitrarily shaped video objects, the MPEG-4 compositor allows the rendering of explicit head models enabling the usage of model-based coding techniques as an alternative for communication with devices connected via low-bandwidth channels. The concept is therefore scalable in the number of participants and channel bit-rates, and can be extended towards future applications of immersive tele-collaboration as well as auto-stereoscopic single- or multi-user displays. REFERENCES 1. M. Chen, Design of a virtual auditorium, in Proc. of ACM Multimedia, (Ottawa, Canada), Sep S. Benford, C. Greenhalgh, and D. Lloyd, Crowded collaborative virtual environments, in Proc. of the ACM SIGCHI Conference on Human Factors in Computing Systems, pp , (Atlanta, USA), Mar O. Ståhl, B. Serenius, and G. Taxén, Meetings for real - experiences from a series of VR-based project meetings, in Proc. Symposium on Virtual Reality Software and Technology, (London, UK), Dec

11 4. V. Rajan, S. Subramanian, D. Keenan, A. Johnson, D. Sandin, and T. DeFanti, A realistic video avatar system for networked virtual environments, in Proc. Immersive Projection Technology Symposium, (Orlando, USA), Mar S. Rauthenberg, A. Graffunder, U. Kowalik, and P. Kauff, Virtual shop and virtual meeting point - two prototype applications of interactive services using the new multimedia coding standard MPEG-4, in Proc. International Conference on Computer Communication, (Tokyo, Japan), Sep W. Chen, H. Towles, L. Nyland, G. Welch, and H. Fuchs, Toward a compelling sensation of telepresence: Demonstrating a portal to a distant (static) office, in Proc. Visualization 2000, pp , (Salt Lake City, USA), Oct S. J. Gibbs, C. Arapis, and C. Breiteneder, TELEPORT - Towards immersive copresence, Multimedia Systems, T. Aoki, W. Kustarto, N. Sakamoto, N. Suzuki, K. Saburi, and H. Yasuda, MONJUnoCHIE system: Videoconference system with eye contact for decision making, in Proc. International Workshop on Advanced Image Technology (IWAIT), P. Kauff and O. Schreer, Virtual team user environments - a step from tele-cubicles towards distributed tele-collaboration in mediated workspaces, in Proc. International Conference on Multimedia and Expo (ICME), (Lausanne, Switzerland), Aug ISO/IEC FDIS , Generic Coding of audio-visual objects: (MPEG-4 video), Final Draft International Standard, Document N2502, D. E. Pearson, Developments in model-based video coding, Proceedings of the IEEE 83, pp , June O. Schreer and P. Kauff, An immersive 3D video-conferencing system using shared virtual team user environments, in ACM Collaborative Environments, CVE 2002, (Bonn, Germany), O. Schreer, I. Feldmann, U. Gölz, and P. Kauff, Fast and robust shadow detection in videoconference applications, in Proc. of VIPromCom, 4th EURASIP IEEE International Symposium on Video Processing and Multimedia Communications, (Zadar, Croatia), June O. Schreer, N. Brandenburg, and P. Kauff, Real-time disparity analysis for applications in imersive teleconference scenarios - a comparative study, in Proc. ICIAP 01, (Palermo, Italy), Sep S. Laveau and O. D. Faugeras, 3D scene representation as a collection of images, in Proc. International Conference on Pattern Recognition (ICPR), pp , (Jerusalem, Israel), Oct S. Avidan and A. Shashua, Novel view synthesis in tensor space, in Proc. Computer Vision and Pattern Recognition, pp , (Puerto Rico), June B. J. Lei and E. A. Hendriks, Multi-step view synthesis with occlusion handling, in Proc. Vision, Modeling, and Visualization VMV 02, (Stuttgart, Germany), Nov W. J. Welsh, S. Searsby, and J. B. Waite, Model-based image coding, British Telecom Technology Journal 8, pp , Jul P. Eisert and B. Girod, Analyzing facial expressions for virtual conferencing, IEEE Computer Graphics and Applications 18, pp , Sep P. Eisert, Very Low Bit-Rate Video Coding Using 3-D Models. PhD thesis, University of Erlangen, Shaker Verlag, Aachen, Germany, P. Eisert and B. Girod, Model-based enhancement of lighting conditions in image sequences, in Proc. SPIE Visual Communications and Image Processing, VCIP-02, (San Jose, USA), Jan